Module 11 Animated and Interactive Graphics

Interactive and animated graphics are one of the major advantages of using the Rmarkdown ecosystem - because you can easily create web pages in markdown (without the pain of HTML), you aren’t limited by paper any more. We’ll cover two different technologies that allow you to create different types of interactive charts, graphs, and interfaces.

It is helpful to think about interactivity in a couple of different ways:

  1. What does it require? Do you need to be doing statistical calculations in the background, or can you precompute all of the data ahead of time?

  2. What type of activity or interactivity do you need?
    • Zoom in/out?
    • Provide additional information in response to user actions (mouseover, click)
    • Provide information over time (animation)
    • Keep track of a data point over multiple plots? (linked plots)
    • Keep track of one or more data points and change their appearance based on user interaction (brushing)
    • Allow the user to change the underlying statistical model or data?

(This is not a full list of all of the types of interactivity, just a few of the more common options)

In this section, we’ll cover two ways to easily create interactive graphics or applets in R. There are, of course, many others – many javascript libraries have R extensions of one form or another.

Module Objectives

  • Create interactive charts with appropriate tools
  • Use Shiny to create interactive web applets

11.1 Plotly

Plotly is a graphing library that uses javascript to add interactivity to graphics. There are several different ways to create plotly graphs in R, but by far the easiest is ggplotly, which converts a ggplot to a plotly plot automatically (so you don’t have to specify most of the details).

11.1.1 ggplotly: ggplot2 to plotly conversions

Set up the data

Let’s try out plotly while doing a bit of exploratory data analysis on this dataset.

Cleaning up volcano

Let’s start by seeing whether the elevation of a volcano changes based on the type of zone it’s on - we might expect that Rift zone volcanos (where plates are pulling away from each other) might not be as high.

But it doesn’t really look like there’s much difference.

Does volcano type makes a difference?

Here, the interactivity actually helps a bit: we don’t need to use the legend to see what each curve corresponds to. We can see that submarine volcanoes are typically much lower in elevation (ok, duh), but also that subglacial volcanoes are found in a very limited range. If we double-click on a legend entry, we can get rid of all other curves and examine each curve one by one.

I added the rug layer after the initial bout because I was curious how much data each of these curves were based on. If we want only curves with n > 10 observations, we can do that:

If we want to specify additional information that should show up in the tooltip, we can do that as well by adding the text aesthetic even though geom_rug doesn’t take a text aesthetic. You may notice that ggplot2 complains about the unknown aesthetic I’ve added to geom_rug: That allows us to mouse over each data point in the rug plot and see what volcano it belongs to. So we can tell from the rug plot that the tallest volcano is Ojas de Salvado, in Chile/Argentina (I believe that translates to Eyes of Salvation?).

At any rate, there isn’t nearly as much variation as I was expecting in the elevation of different types of volcanoes.

ggplotly makes it very easy to generate plots that have a ggplot2 equivalent; you can customize these plots further using plotly functions that we’ll see in the next section. But first, try the interface out on your own.

Try it out

Conduct an exploratory data analysis of the eruptions dataset. What do you find?

My solution


head(eruptions)
# A tibble: 6 x 15
  volcano_number volcano_name eruption_number eruption_catego… area_of_activity
           <dbl> <chr>                  <dbl> <chr>            <chr>           
1         266030 Soputan                22354 Confirmed Erupt… <NA>            
2         343100 San Miguel             22355 Confirmed Erupt… <NA>            
3         233020 Fournaise, …           22343 Confirmed Erupt… <NA>            
4         345020 Rincon de l…           22346 Confirmed Erupt… <NA>            
5         353010 Fernandina             22347 Confirmed Erupt… <NA>            
6         273070 Taal                   22344 Confirmed Erupt… <NA>            
# … with 10 more variables: vei <dbl>, start_year <dbl>, start_month <dbl>,
#   start_day <dbl>, evidence_method_dating <chr>, end_year <dbl>,
#   end_month <dbl>, end_day <dbl>, latitude <dbl>, longitude <dbl>

summary(eruptions %>% mutate(eruption_category = factor(eruption_category)))
 volcano_number   volcano_name       eruption_number
 Min.   :210010   Length:11178       Min.   :10001  
 1st Qu.:263310   Class :character   1st Qu.:12817  
 Median :290050   Mode  :character   Median :15650  
 Mean   :300284                      Mean   :15667  
 3rd Qu.:343030                      3rd Qu.:18464  
 Max.   :600000                      Max.   :22355  
                                                    
            eruption_category area_of_activity        vei        
 Confirmed Eruption  :9900    Length:11178       Min.   :0.0000  
 Discredited Eruption: 166    Class :character   1st Qu.:1.0000  
 Uncertain Eruption  :1112    Mode  :character   Median :2.0000  
                                                 Mean   :1.9479  
                                                 3rd Qu.:2.0000  
                                                 Max.   :7.0000  
                                                 NA's   :2906    
   start_year         start_month        start_day      evidence_method_dating
 Min.   :-11345.00   Min.   : 0.0000   Min.   : 0.000   Length:11178          
 1st Qu.:   680.00   1st Qu.: 0.0000   1st Qu.: 0.000   Class :character      
 Median :  1847.00   Median : 1.0000   Median : 0.000   Mode  :character      
 Mean   :   622.85   Mean   : 3.4509   Mean   : 7.015                         
 3rd Qu.:  1950.00   3rd Qu.: 7.0000   3rd Qu.:15.000                         
 Max.   :  2020.00   Max.   :12.0000   Max.   :31.000                         
 NA's   :1           NA's   :193       NA's   :196                            
    end_year        end_month          end_day          latitude      
 Min.   :-475.0   Min.   : 0.0000   Min.   : 0.000   Min.   :-77.530  
 1st Qu.:1895.0   1st Qu.: 3.0000   1st Qu.: 4.000   1st Qu.: -6.102  
 Median :1957.0   Median : 6.0000   Median :15.000   Median : 17.600  
 Mean   :1917.3   Mean   : 6.2208   Mean   :13.316   Mean   : 16.866  
 3rd Qu.:1992.0   3rd Qu.: 9.0000   3rd Qu.:21.000   3rd Qu.: 40.821  
 Max.   :2020.0   Max.   :12.0000   Max.   :31.000   Max.   : 85.608  
 NA's   :6846     NA's   :6849      NA's   :6852                      
   longitude       
 Min.   :-179.970  
 1st Qu.: -77.656  
 Median :  55.708  
 Mean   :  31.572  
 3rd Qu.: 139.394  
 Max.   : 179.580  
                   


# Historical (very historical) dates are a bit of a pain to work with, so I 
# wrote a helper function which takes year, month, and day arguments and formats
# them properly

fix_date <- function(yyyy, mm, dd) {
  # First, negative years (BCE) are a bit of a problem.
  neg <- yyyy < 0
  subtract_years <- pmax(-yyyy, 0) # Years to subtract off later
  # for now, set to 0
  year_fixed <- pmax(yyyy, 0) # this will set anything negative to 0
  
  # sometimes the day or month isn't known, so just use 1 for both. 
  # recorded value may be NA or 0.
  day_fixed <- ifelse(is.na(dd), 1, pmax(dd, 1))
  month_fixed <- ifelse(is.na(mm), 1, pmax(mm, 1))
  
  # Need to format things precisely, so use sprintf
  # %0xd ensures that you have at least x digits, padding the left side with 0s
  # lubridate doesn't love having 3-digit years. 
  date_str <- sprintf("%04d/%02d/%02d", year_fixed, month_fixed, day_fixed)
  # Then we can convert the dates and subtract off the years for pre-CE dates
  date <- ymd(date_str) - years(subtract_years)
}

erupt <- eruptions %>% 
  # Don't work with discredited eruptions
  filter(eruption_category == "Confirmed Eruption") %>%
  # Create start and end dates
  mutate(
    start_date = fix_date(start_year, start_month, start_day),
    end_date = fix_date(end_year, end_month, end_day),
    # To get duration, we have to start with a time interval, 
    # convert to duration, then convert to a numeric value
    duration = interval(start = start_date, end = end_date) %>% 
      as.duration() %>% 
      as.numeric("days"))
Warning: 1 failed to parse.
Warning: 5895 failed to parse.

Let’s start out seeing what month most eruptions occur in…

Another numerical variable is VEI, volcano explosivity index. A VEI of 0 is non-explosive, a VEI of 4 is about what Mt. St. Helens hit in 1980, and a VEI of 5 is equivalent to the Krakatau explosion in 1883. A VEI of 8 would correspond to a major Yellowstone caldera eruption (which hasn’t happened for 600,000 years). Basically, VEI increase of 1 is an order of magnitude change in the amount of material the eruption released.

We can also look at the frequency of eruptions over time. We’ll expect some historical bias - we don’t have exact dates for some of these eruptions, and if no one was around to write the eruption down (or the records were destroyed) there’s not going to be a date listed here.

As expected, it’s pretty rare to see many eruptions before ~1800 AD, which is about when we have reliable historical records41 for most of the world (exceptions include e.g. Vestuvius, which we have extensive written information about).

It seems that the really big eruptions might be less likely to last for a long time, but it is hard to tell because there aren’t that many of them (thankfully).

11.1.2 plot_ly: Like base plotting, but interactive!

You can also create plotly charts that aren’t limited by what you can do in ggplot2, using the plot_ly function.

Plotly cheat sheet

We can start with a scatterplot of volcanoes along the Earth’s surface:

And then we can start customizing:

The plot_ly function is also pipe friendly. Variable mappings are preceded with ~ to indicate that the visual appearance changes with the value of the variable.

Plotly will handle some variable mappings for you, depending on which “trace” type (plot/geom) you’re using.

The plotly documentation often uses plyr – which is an older version of dplyr. If you load plyr, it will seriously mess up your day – a lot of the function names are the same. So, instead, here’s a shortcut: spread is pivot_wider and gather is pivot_longer. That should at least help with understanding what the code is doing.

If you do accidentally load plyr, that’s fine: just restart your R session so that your loaded packages are cleared and you can start over. Or, if you must, you can reference a plyr function using plyr::spread without loading the package – that’s a safe way to use the plotly demo code as-is.

Let’s explore traces a bit. According to the plotly documentation,

A trace is just the name we give a collection of data and the specifications of which we want that data plotted. Notice that a trace will also be an object itself, and these will be named according to how you want the data displayed on the plotting surface

In ggplot2 terms, it seems that a trace is somewhat akin to a geom.

But, if you want all of the variables to be shown with the same trace type, it’s probably easier to get to long form:

There are many different trace types in plotly, but your best bet is to check the documentation to see what is available.

11.1.3 Animation

Plotly can also animate your plots for you.

Sometimes the animations get a bit trippy, don’t they?

You can even animate by something other than time, if you’re so inclined…

There are other types of animations as well, including the ability to change plot formats, trace types, and more.

11.2 Leaflet maps

Leaflet is another javascript library that allows for interactive data visualization. We’re only going to briefly talk about it here, but there is extensive documentation that includes details of how to work with different types of geographical data, chloropleth maps, plugins, and more.

To explore the leaflet package, we’ll start out playing with a dataset of Bigfoot sightings assembled from the Bigfoot Field Researchers Organization’s Google earth tool

Parsed with column specification:
cols(
  .default = col_double(),
  observed = col_character(),
  location_details = col_character(),
  county = col_character(),
  state = col_character(),
  season = col_character(),
  title = col_character(),
  date = col_date(format = ""),
  classification = col_character(),
  geohash = col_character(),
  precip_type = col_character(),
  summary = col_character()
)
See spec(...) for full column specifications.

We can start out by plotting a map with the location of each sighting. I’ve colored the points in a seasonal color scheme, and added the description of each incident as a mouseover label.

Of course, because this is an interactive map library, we aren’t limited to any one scale. We can also plot data at the city level:

We can also plot regions, instead of just points. I downloaded a dataset released by the US Forest Service, Bailey’s Ecoregions and Subregions dataset, which categorizes the US into different climate and ecological zones.

To map colors to variables, we have to define a color palette and variable mapping ourselves, and pass that function into the leaflet object we’re adding.

Try it out

Download the Shapefiles for the 116th Congress Congressional Districts. Unzip the file and read it in using the code below (you’ll have to change the file path).

Use the MIT Election Data and Science Lab’s US House election results42, and merge this data with the shapefiles to plot the results of the 2018 midterms in a way that you think is useful (you can use any of the available data).

Some notes: - FIPS codes are used to identify the state and district, with 00 indicating at-large districts (one district for the state) and 98 indicating non-voting districts. - If you would like to add in the number of citizens of voting age, you can get that information here but you will have to do some cleaning in order to join the table with the others. - Minnesota’s Democratic-farmer-labor party caucuses with the Democrats but maintains its name for historical reasons. You can safely recode this if you want to.

One solution

library(sf)
library(htmltools) # to mark labels as html code

# Read in the results
election_results <- read_csv("data/1976-2018-house2.csv") %>%
  filter(year == 2018) %>%
  mutate(state_fips = sprintf("%02d", state_fips), 
         district = sprintf("%02d", district)) %>%
  group_by(state, state_fips, state_po, district, stage) %>%
  arrange(candidatevotes) %>%
  mutate(pct = candidatevotes/totalvotes) %>%
  # Keep the winner only
  filter(pct == max(pct)) %>%
  # Fix Minnesota
  mutate(party = ifelse(party == "democratic-farmer-labor", "democrat", party))
Parsed with column specification:
cols(
  year = col_double(),
  state = col_character(),
  state_po = col_character(),
  state_fips = col_double(),
  state_cen = col_double(),
  state_ic = col_double(),
  office = col_character(),
  district = col_double(),
  stage = col_character(),
  runoff = col_logical(),
  special = col_logical(),
  candidate = col_character(),
  party = col_character(),
  writein = col_logical(),
  mode = col_character(),
  candidatevotes = col_double(),
  totalvotes = col_double(),
  unofficial = col_logical(),
  version = col_double()
)

# Read in the districts
congress_districts <- st_read("data/116_congress/cb_2018_us_cd116_5m.shp")
Reading layer `cb_2018_us_cd116_5m' from data source `/home/susan/Projects/Class/unl-stat850/2020-stat850/data/116_congress/cb_2018_us_cd116_5m.shp' using driver `ESRI Shapefile'
Simple feature collection with 441 features and 8 fields
geometry type:  MULTIPOLYGON
dimension:      XY
bbox:           xmin: -179.14734 ymin: -14.552549 xmax: 179.77847 ymax: 71.352561
CRS:            4269

# Clean up congress districts
congress_districts <- congress_districts %>%
  # Convert factors to characters
  mutate(across(where(is.factor), as.character)) %>%
  # Handle at-large districts
  mutate(district = ifelse(CD116FP == "00", "01", CD116FP))

# Merge
congress_districts <- congress_districts %>% 
  left_join(election_results, by = c("STATEFP" = "state_fips", "CD116FP" = "district")) %>%
  mutate(party = factor(party, levels = c("republican", "democrat")),
         short_party = ifelse(party == "republican", "R", "D"),
         label = paste0(state_po, "-", district, candidate, " (", short_party, ")"))

# Define a palette
region_pal <- colorFactor(c("#e9141d", "#0015bc"), congress_districts$party)

congress_districts %>%
leaflet() %>%
  addTiles() %>%
  addPolygons(stroke = TRUE, fillOpacity = ~pct/2, 
              # still want to see what's underneath, even in safe districts
              fillColor = ~region_pal(party), color = ~region_pal(party), 
              label = ~label)
Warning: sf layer has inconsistent datum (+proj=longlat +datum=NAD83 +no_defs ).
Need '+proj=longlat +datum=WGS84'

11.3 Shiny

https://github.com/rfordatascience/tidytuesday/blob/master/data/2020/2020-05-26/readme.md

Before we get started on Shiny, take a few minutes and poke around the RStudio Shiny user showcase. Sometimes, it helps to have some motivation, and to get a sense of what is possible before you start learning something. One of the more amusing ones I found was an exploration of lego demographics.

All done?

Shiny is a framework for building interactive web applications in R. Unlike plotly and other graphics engines, Shiny depends on an R instance on a server to do computations. This means Shiny is much more powerful and has more capabilities, but also that it’s harder to share and deploy - you have to have access to a web server with R installed on it. If you happen to have a server like that, though, Shiny is pretty awesome. RStudio runs a service called shinyapps.io that will provide some limited free hosting, as well as paid plans for apps that have more web traffic, but you can also create Shiny apps for local use - I often do this for model debugging when I’m using neural networks, because they’re so complicated.

RStudio has a set of well produced video tutorials to introduce Shiny. I’d recommend you at least listen to the introduction if you’re a visual/audio learner (the whole tutorial is about 2 hours long). There is also a written tutorial if you prefer to learn in written form (7 lessons, each is about 20 minutes long).

I generally think it’s better to send you to the source when there are well-produced resources, rather than trying to rehash something to put my own spin on it.

One other interesting feature to keep in mind when using Shiny - you can integrate Shiny reactivity into Rmarkdown by adding runtime: shiny to the markdown header.

11.4 General References

11.4.1 Other interactive tools

  • htmlwidgets - a generic wrapper for any Javascript library (htmlwidgets is used under the hood in both Leaflet and Plotly R integration)

  • dash - Another dashboard program supported by plotly. dash is the python equivalent of shiny, but also has R integration (though I’m not sure how well it’s supported).

  1. There are obviously exceptions - we can figure out the exact date and approximate time that there was an earthquake along the Cascadia subduction zone based on a combination of oral histories of the indigenous people and records of a massive tsunami in Japan Excellent read, if you’re interested, and the Nature paper.↩

  2. Alternately, you can find the rehosted file here.↩